[AArch64] analyse dependences of ldp/stp

[mcrosier]

Summary

Add load and store pairs to the dependence analysis of offset and width.
This fixes PR26358.

Patch by Sebastian Pop and Abderrazek Zaafrani.

[qcolombet]

llvm/lib/Target/AArch64/AArch64InstrInfo.cpp

1446

s/else if/else/
And move the condition of the if in an assert.

llvm/test/CodeGen/AArch64/ldst-pairing.ll

3

CHECK-LABEL:

+ Give a description of what is tested.

20

ditto

48

ditto

[sebpop]

Address comments from Quentin.

[qcolombet]

llvm/test/CodeGen/AArch64/ldst-pairing.ll

5

Could you be more specific?
In particular, explain what is the difference between @st1, @st2, and so on.
Put differently what is relevant in this test that was not previously tested.

For instance, the FIXME in st5 is a good example. We check the merging of stores when the access are not in increasing order of the address location.

[mcrosier]

Please add the non-temporal ldp/stp instructions. This should probably land after D17097.

[sebpop]

Add nontemporal, update testcase comments.

[mcrosier]

All of the provided tests pass on current top of trunk (i.e., without this patch). We'll need some tests that exercise the patch before it can be committed.

llvm/lib/Target/AArch64/AArch64InstrInfo.cpp

1346

I think the 'or' is misleading as we're looking for a base register followed by an immediate offset in both the non-paired and paired cases.

We're looking for 'ldr x1, [x0, #imm]' or 'ldp x1, x2, [x0, #imm]'

Therefore, I'd suggest leaving this comment as is and add comments below.

1348

// Non-paired instruction (e.g., ldr x1, [x0, #8]).

1351

// Paired instruction (e.g., ldp x1, x2, [x0, #8]).

[mcrosier]

llvm/test/CodeGen/AArch64/ldst-pairing.ll

2

You may need to disable the MI scheduler for this to work. I believe you can do this by specifying -disable-post-ra on the llc command line.

[mcrosier]

LGTM, assuming there are no correctness issues.

Please only commit the st2 test as the other tests are either not related to the patch our are redundant with st2. Also, the test can be included in arm64-stp-aa.ll, rather than creating a new file.

llvm/test/CodeGen/AArch64/ldst-pairing.ll

9

This tests the most basic pairing capabilities and is not relevant to the patch at hand (i.e., this passes without this patch).

31

This is the most relevant test to the commit. To properly test the patch the pre-RA MI scheduler must be disabled (-enable-misched=false) and D17097 must also be applied.

63

Same story at st1. This tests the most basic pairing capabilities and is not relevant to the patch at hand (i.e., this passes without this patch).

82

This is redundant and not needed.

104

This test is not relevant to this patch. It can be committed separately with a FIXME comment in AArch64LoadStoreOptimizer.cpp.

[mcrosier]

I wanted to push this patch along, so I ran some correctness tests. It turns out the paired instructions aren't being scaled properly (see inline comments). After addressing those issues I'm now seeing a few regressions due to fewer stp instructions. I'm going to block this patch while I investigate.

llvm/lib/Target/AArch64/AArch64InstrInfo.cpp

1405

Scale = 16;
Width = 32;

1414

Scale = 8;
Width = 16;

1420

Scale = 4;
Width = 8;

[mcrosier]

PR26879 may be part of the problem.
https://llvm.org/bugs/show_bug.cgi?id=26879

[mcrosier]

Correct scale/offset issues and address all previous feedback.

[sebpop]

In D17098#370005, @mcrosier wrote:

Correct scale/offset issues and address all previous feedback.

The corrected version looks very similar to what I remember having implemented first, and then we found out that the scale did not match.
Let's not apply this patch yet: I need to discuss this again with my colleague Abderrazek and we will try to find why we changed to scale = width for ldp/stp formation. We will come back with a counter-example if we find one.

[mcrosier]

In D17098#370320, @sebpop wrote:

In D17098#370005, @mcrosier wrote:

Correct scale/offset issues and address all previous feedback.

The corrected version looks very similar to what I remember having implemented first, and then we found out that the scale did not match.
Let's not apply this patch yet: I need to discuss this again with my colleague Abderrazek and we will try to find why we changed to scale = width for ldp/stp formation. We will come back with a counter-example if we find one.

Sounds good. Let me know if I've missed something. The matching scale and offsets caused a large number of correctness failures in the test-suite as well as SPEC200X in my testing.

[sebpop]

In D17098#370340, @mcrosier wrote:

Let me know if I've missed something. The matching scale and offsets caused a large number of correctness failures in the test-suite as well as SPEC200X in my testing.

Chad, your changes to the patch are correct: the ARM documentation states that in the case of a pair of 32bit regs the immediate byte offset is a multiple of 4.
The patch LGTM.

[sebpop]

Ping.
Abderrazek has looked at the code produced with the patch on xalan benchmark
and he found that with the patch there are 13 more stp, and not fewer stp.
Ok to commit?

[sebpop]

Chad, is it ok to commit this patch?

[junbuml]

Please rebase it. LGTM.

lib/Target/AArch64/AArch64InstrInfo.cpp

1351 ↗

(On Diff #50064)

Please use getNumExplicitOperands().

[mcrosier]

In D17098#402547, @junbuml wrote:

Please rebase it. LGTM.

Approved per Jun and Sebastian. Jun verified the regression I was seeing no longer exists. I'll go ahead and rebase the patch and commit with Jun's suggestion to use getNumExplicitOperands(). Thanks, Jun/Sebastian.

[mcrosier]

Committed in r266462.

llvm/lib/Target/AArch64/AArch64InstrInfo.cpp

case AArch64::LDRWui:
     Offset = LdSt->getOperand(2).getImm() * Width;
     return true;
   };
 }
 
 bool AArch64InstrInfo::getMemOpBaseRegImmOfsWidth(
     MachineInstr *LdSt, unsigned &BaseReg, int &Offset, int &Width,
     const TargetRegisterInfo *TRI) const {
-  // Handle only loads/stores with base register followed by immediate offset.
-  if (LdSt->getNumOperands() != 3)
-    return false;
+  // Handle only loads/stores with base register followed by immediate offset
+  // or pairs.
+  if (LdSt->getNumOperands() == 3) {
     if (!LdSt->getOperand(1).isReg() || !LdSt->getOperand(2).isImm())
         return false;
+  } else if (LdSt->getNumOperands() == 4) {
+    if (!LdSt->getOperand(1).isReg() || !LdSt->getOperand(2).isReg() || !LdSt->getOperand(3).isImm())
+      return false;
+  } else {
+    return false;
+  }
 
   // Offset is calculated as the immediate operand multiplied by the scaling factor.
   // Unscaled instructions have scaling factor set to 1.
   int Scale = 0;
   switch (LdSt->getOpcode()) {
   default:
     return false;
   case AArch64::LDURQi:
bool AArch64InstrInfo::getMemOpBaseRegImmOfsWidth(
   case AArch64::LDURBBi:
   case AArch64::LDURSBXi:
   case AArch64::LDURSBWi:
   case AArch64::STURBi:
   case AArch64::STURBBi:
     Width = 1;
     Scale = 1;
     break;
+  case AArch64::LDPQi:
+  case AArch64::STPQi:
+    Scale = Width = 32;
+    break;
+  case AArch64::LDPXi:
+  case AArch64::LDPDi:
+  case AArch64::STPXi:
+  case AArch64::STPDi:
   case AArch64::LDRQui:
   case AArch64::STRQui:
     Scale = Width = 16;
     break;
+  case AArch64::LDPWi:
+  case AArch64::STPWi:
+  case AArch64::LDPSi:
+  case AArch64::STPSi:
   case AArch64::LDRXui:
   case AArch64::LDRDui:
   case AArch64::STRXui:
   case AArch64::STRDui:
     Scale = Width = 8;
     break;
   case AArch64::LDRWui:
   case AArch64::LDRSui:
   case AArch64::STRWui:
   case AArch64::STRSui:
     Scale = Width = 4;
     break;
   case AArch64::LDRHui:
   case AArch64::LDRHHui:
   case AArch64::STRHui:
   case AArch64::STRHHui:
     Scale = Width = 2;
     break;
   case AArch64::LDRBui:
   case AArch64::LDRBBui:
   case AArch64::STRBui:
   case AArch64::STRBBui:
     Scale = Width = 1;
     break;
   }
 
+  if (LdSt->getNumOperands() == 3) {
     BaseReg = LdSt->getOperand(1).getReg();
     Offset = LdSt->getOperand(2).getImm() * Scale;
+  } else if (LdSt->getNumOperands() == 4) {
+    BaseReg = LdSt->getOperand(2).getReg();
+    Offset = LdSt->getOperand(3).getImm() * Scale;
+  }
   return true;
 }
 
 /// Detect opportunities for ldp/stp formation.
 ///
 /// Only called for LdSt for which getMemOpBaseRegImmOfs returns true.
 bool AArch64InstrInfo::shouldClusterLoads(MachineInstr *FirstLdSt,
                                           MachineInstr *SecondLdSt,

llvm/test/CodeGen/AArch64/aarch64-a57-fp-load-balancing.ll

entry:
   store double %add19, double* %arrayidx20, align 8
   ret void
 }
 
 ; Overlapping groups - coloring needed.
 
 ; CHECK-LABEL: f2:
 ; CHECK-EVEN: fmadd [[x:d[0-9]*[02468]]]
-; CHECK-EVEN: fmul [[y:d[0-9]*[13579]]]
 ; CHECK-ODD: fmadd [[x:d[0-9]*[13579]]]
+; CHECK-A57: fmadd [[x]]
+; CHECK-EVEN: fmul [[y:d[0-9]*[13579]]]
 ; CHECK-ODD: fmul [[y:d[0-9]*[02468]]]
-; CHECK: fmadd [[x]]
 ; CHECK: fmadd [[y]]
+; CHECK-A53: fmadd [[x]]
+; CHECK-A53: fmadd [[y]]
 ; CHECK: fmsub [[x]]
-; CHECK: fmadd [[y]]
+; CHECK-A57: fmadd [[y]]
+; CHECK-A53-DAG: str [[y]]
 ; CHECK: fmadd [[x]]
 ; CHECK-A57: stp [[x]], [[y]]
 ; CHECK-A53-DAG: str [[x]]
-; CHECK-A53-DAG: str [[y]]
 
 define void @f2(double* nocapture readonly %p, double* nocapture %q) #0 {
 entry:
   %0 = load double, double* %p, align 8
   %arrayidx1 = getelementptr inbounds double, double* %p, i64 1
   %1 = load double, double* %arrayidx1, align 8
   %arrayidx2 = getelementptr inbounds double, double* %p, i64 2
   %2 = load double, double* %arrayidx2, align 8

 }
 
 declare void @g(...) #1
 
 ; Single precision version of f2.
 
 ; CHECK-LABEL: f4:
 ; CHECK-EVEN: fmadd [[x:s[0-9]*[02468]]]
-; CHECK-EVEN: fmul [[y:s[0-9]*[13579]]]
 ; CHECK-ODD: fmadd [[x:s[0-9]*[13579]]]
+; CHECK-A57: fmadd [[x]]
+; CHECK-EVEN: fmul [[y:s[0-9]*[13579]]]
 ; CHECK-ODD: fmul [[y:s[0-9]*[02468]]]
-; CHECK: fmadd [[x]]
 ; CHECK: fmadd [[y]]
+; CHECK-A53: fmadd [[x]]
+; CHECK-A53: fmadd [[y]]
 ; CHECK: fmsub [[x]]
-; CHECK: fmadd [[y]]
+; CHECK-A57: fmadd [[y]]
+; CHECK-A53-DAG: str [[y]]
 ; CHECK: fmadd [[x]]
 ; CHECK-A57: stp [[x]], [[y]]
 ; CHECK-A53-DAG: str [[x]]
-; CHECK-A53-DAG: str [[y]]
 
 define void @f4(float* nocapture readonly %p, float* nocapture %q) #0 {
 entry:
   %0 = load float, float* %p, align 4
   %arrayidx1 = getelementptr inbounds float, float* %p, i64 1
   %1 = load float, float* %arrayidx1, align 4
   %arrayidx2 = getelementptr inbounds float, float* %p, i64 2
   %2 = load float, float* %arrayidx2, align 4

llvm/test/CodeGen/AArch64/arm64-alloc-no-stack-realign.ll

 ; RUN: llc < %s -mtriple=arm64-apple-darwin -enable-misched=false | FileCheck %s
 
 ; rdar://12713765
 ; Make sure we are not creating stack objects that are assumed to be 64-byte
 ; aligned.
 @T3_retval = common global <16 x float> zeroinitializer, align 16
 
 define void @test(<16 x float>* noalias sret %agg.result) nounwind ssp {
 entry:
 ; CHECK: test
-; CHECK: stp [[Q1:q[0-9]+]], [[Q2:q[0-9]+]], [sp, #32]
 ; CHECK: stp [[Q1:q[0-9]+]], [[Q2:q[0-9]+]], [sp]
+; CHECK: stp [[Q1:q[0-9]+]], [[Q2:q[0-9]+]], [sp, #32]
 ; CHECK: stp [[Q1:q[0-9]+]], [[Q2:q[0-9]+]], {{\[}}[[BASE:x[0-9]+]], #32]
 ; CHECK: stp [[Q1:q[0-9]+]], [[Q2:q[0-9]+]], {{\[}}[[BASE]]]
  %retval = alloca <16 x float>, align 16
  %0 = load <16 x float>, <16 x float>* @T3_retval, align 16
  store <16 x float> %0, <16 x float>* %retval
  %1 = load <16 x float>, <16 x float>* %retval
  store <16 x float> %1, <16 x float>* %agg.result, align 16
  ret void
 }

llvm/test/CodeGen/AArch64/fastcc.ll

 ; CHECK-TAIL: str w{{[0-9]+}}, [sp, #-32]!
 
 
   call fastcc void @func_stack8([8 x i32] undef, i32 42)
 ; CHECK:  bl func_stack8
 ; CHECK-NOT: sub sp, sp,
 
 ; CHECK-TAIL: bl func_stack8
-; CHECK-TAIL: sub sp, sp, #16
+; CHECK-TAIL: stp xzr, xzr, [sp, #-16]!
 
 
   call fastcc void @func_stack32([8 x i32] undef, i128 0, i128 9)
 ; CHECK: bl func_stack32
 ; CHECK-NOT: sub sp, sp,
 
 
 ; CHECK-TAIL: bl func_stack32
; CHECK-TAIL: str w{{[0-9]+}}, [sp, #-32]!
 
 
   call fastcc void @func_stack8([8 x i32] undef, i32 42)
 ; CHECK:  bl func_stack8
 ; CHECK-NOT: sub sp, sp,
 
 
 ; CHECK-TAIL: bl func_stack8
-; CHECK-TAIL: sub sp, sp, #16
+; CHECK-TAIL: stp xzr, xzr, [sp, #-16]!
 
 
   call fastcc void @func_stack32([8 x i32] undef, i128 0, i128 9)
 ; CHECK: bl func_stack32
 ; CHECK-NOT: sub sp, sp,
 
 
 ; CHECK-TAIL: bl func_stack32

 ; CHECK-TAIL: mov x29, sp
 
 
   call fastcc void @func_stack8([8 x i32] undef, i32 42)
 ; CHECK:  bl func_stack8
 ; CHECK-NOT: sub sp, sp,
 
 ; CHECK-TAIL: bl func_stack8
-; CHECK-TAIL: sub sp, sp, #16
+; CHECK-TAIL: stp xzr, xzr, [sp, #-16]!
 
 
   call fastcc void @func_stack32([8 x i32] undef, i128 0, i128 9)
 ; CHECK: bl func_stack32
 ; CHECK-NOT: sub sp, sp,
 
 
 ; CHECK-TAIL: bl func_stack32

llvm/test/CodeGen/AArch64/ldst-pairing.ll

+; RUN: llc < %s -march=arm64 -mtriple=aarch64-none-linux-gnu | FileCheck %s
+
+; CHECK: stp
+; CHECK: stp
+define void @st1(<4 x float> %a, <4 x float> %b, <4 x float> %c, <4 x float> %d, <4 x float> * %base, i64 %index) {
+entry:
+  %a1 = getelementptr inbounds <4 x float>, <4 x float>* %base, i64 %index
+  %b1 = getelementptr <4 x float>, <4 x float>* %a1, i64 1
+  %c1 = getelementptr <4 x float>, <4 x float>* %a1, i64 2
+  %d1 = getelementptr <4 x float>, <4 x float>* %a1, i64 3
+
+  store <4 x float> %a, <4 x float> * %a1
+  store <4 x float> %b, <4 x float> * %b1
+  store <4 x float> %c, <4 x float> * %c1
+  store <4 x float> %d, <4 x float> * %d1
+
+  ret void
+}
+
+; CHECK: stp
+; CHECK: stp
+define void @st2(<4 x float> %a, <4 x float> %b, <4 x float> %c, <4 x float> %d, float* %base, i64 %index) {
+entry:
+  %a0 = getelementptr inbounds float, float* %base, i64 %index
+  %b0 = getelementptr float, float* %a0, i64 4
+  %c0 = getelementptr float, float* %a0, i64 8
+  %d0 = getelementptr float, float* %a0, i64 12
+
+  %a1 = bitcast float* %a0 to <4 x float>*
+  %b1 = bitcast float* %b0 to <4 x float>*
+  %c1 = bitcast float* %c0 to <4 x float>*
+  %d1 = bitcast float* %d0 to <4 x float>*
+
+  store <4 x float> %c, <4 x float> * %c1, align 4
+  store <4 x float> %a, <4 x float> * %a1, align 4
+
+  ; This fadd forces the compiler to pair %c and %e after fadd, and leave the
+  ; stores %a and %b separated by a stp. The dependence analysis needs then to
+  ; prove that it is safe to move %b past the stp to be paired with %a.
+  %e = fadd fast <4 x float> %d, %a
+
+  store <4 x float> %e, <4 x float> * %d1, align 4
+  store <4 x float> %b, <4 x float> * %b1, align 4
+
+  ret void
+}
+
+; CHECK: stp
+; CHECK: stp
+define void @st3(<4 x float> %a, <4 x float> %b, <4 x float> %c, <4 x float> %d, <4 x float> * %base, i64 %index) {
+entry:
+  %a1 = getelementptr inbounds <4 x float>, <4 x float>* %base, i64 %index
+  %b1 = getelementptr <4 x float>, <4 x float>* %a1, i64 1
+  %c1 = getelementptr <4 x float>, <4 x float>* %a1, i64 2
+  %d1 = getelementptr <4 x float>, <4 x float>* %a1, i64 3
+
+  store <4 x float> %a, <4 x float> * %a1
+  store <4 x float> %d, <4 x float> * %d1
+  store <4 x float> %c, <4 x float> * %c1
+  store <4 x float> %b, <4 x float> * %b1
+
+  ret void
+}
+
+; CHECK: stp
+; CHECK: stp
+define void @st4(<4 x float> %a, <4 x float> %b, <4 x float> %c, <4 x float> %d, <4 x float> * %base, i64 %index) {
+entry:
+  %a1 = getelementptr inbounds <4 x float>, <4 x float>* %base, i64 %index
+  %b1 = getelementptr <4 x float>, <4 x float>* %a1, i64 1
+  %c1 = getelementptr <4 x float>, <4 x float>* %a1, i64 2
+  %d1 = getelementptr <4 x float>, <4 x float>* %a1, i64 3
+
+  store <4 x float> %d, <4 x float> * %d1
+  store <4 x float> %b, <4 x float> * %b1
+
+  %e = fadd fast <4 x float> %c, %a
+
+  store <4 x float> %e, <4 x float> * %c1
+  store <4 x float> %a, <4 x float> * %a1
+
+  ret void
+}
+
+; FIXME: st5 should contain two stp. The current pairing algorithm is greedy and
+; is pairing %c and %b, leaving the two other stores %a and %d not pairable.
+
+; CHECK: stp
+define void @st5(<4 x float> %a, <4 x float> %b, <4 x float> %c, <4 x float> %d, <4 x float> * %base, i64 %index) {
+entry:
+  %a1 = getelementptr inbounds <4 x float>, <4 x float>* %base, i64 %index
+  %b1 = getelementptr <4 x float>, <4 x float>* %a1, i64 1
+  %c1 = getelementptr <4 x float>, <4 x float>* %a1, i64 2
+  %d1 = getelementptr <4 x float>, <4 x float>* %a1, i64 3
+
+  store <4 x float> %c, <4 x float> * %c1
+  store <4 x float> %a, <4 x float> * %a1
+  store <4 x float> %b, <4 x float> * %b1
+  store <4 x float> %d, <4 x float> * %d1
+
+  ret void
+}